[Paper Review] Causal Tree Estimation of Heterogeneous Household Response to Time-Of-Use Electricity Pricing Schemes

Eoghan O’Neill, Melvyn Weeks|arXiv (Cornell University)|Oct 22, 2018

Smart Grid Energy Management17 citations

TL;DR

This paper uses causal forest machine learning to estimate heterogeneous household responses to time-of-use (TOU) electricity pricing, revealing that past electricity consumption patterns—especially weekday usage levels and variability—are stronger predictors of demand response than survey variables. Younger, more educated, and higher-consuming households show greater responsiveness, with usage data outperforming demographic information in predicting treatment effects.

ABSTRACT

We examine the household-specific effects of the introduction of Time-of-Use (TOU) electricity pricing schemes. Using a causal forest (Athey and Imbens, 2016; Wager and Athey, 2018; Athey et al., 2019), we consider the association between past consumption and survey variables, and the effect of TOU pricing on household electricity demand. We describe the heterogeneity in household variables across quartiles of estimated demand response and utilise variable importance measures. Household-specific estimates produced by a causal forest exhibit reasonable associations with covariates. For example, households that are younger, more educated, and that consume more electricity, are predicted to respond more to a new pricing scheme. In addition, variable importance measures suggest that some aspects of past consumption information may be more useful than survey information in producing these estimates.

Motivation & Objective

• To estimate heterogeneous treatment effects of time-of-use (TOU) electricity pricing on household demand using machine learning.
• To identify which household characteristics—demographic or consumption-based—best predict responsiveness to TOU tariffs.
• To assess the relative importance of survey data versus historical electricity consumption data in predicting individual-level demand response.
• To address variable selection bias in tree-based models by using permutation-based variable importance tests.
• To provide policymakers with insights into which households are most likely to respond to TOU pricing, informing targeted energy policy design.

Proposed method

• Employs causal forest (a random forest of causal trees) to estimate individual-level treatment effects using potential outcomes framework.
• Uses conditional average treatment effect (CATE) estimation to model heterogeneous responses across household subpopulations defined by covariates.
• Applies variable importance measures based on split frequency across trees, with depth-weighted aggregation to rank predictors.
• Implements permutation-based tests to correct for bias toward continuous variables with more splitting points in variable importance assessment.
• Fits separate causal forests using only survey data, only usage data, and combined data to compare predictive performance.
• Uses quantile-based discretization of continuous usage variables to test robustness of variable importance results.

Experimental results

Research questions

• RQ1 Which household characteristics are most predictive of responsiveness to time-of-use electricity pricing?
• RQ2 How does the predictive power of electricity consumption data compare to that of survey-based demographic variables in estimating demand response?
• RQ3 To what extent is heterogeneity in response driven by past consumption behavior versus socioeconomic factors?
• RQ4 Are variable importance measures in causal forests biased toward continuous usage variables, and how can this be corrected?
• RQ5 Can the causal forest model identify distinct subgroups of households with markedly different responses to TOU pricing?

Key findings

• Households that are younger, more educated, and consume more electricity exhibit significantly greater demand response to TOU pricing.
• Variable importance measures rank past electricity usage—particularly weekday consumption levels and variance—highest, indicating they are the most informative predictors of response.
• Permutation-based tests confirm that usage variables are more important than survey variables, even after correcting for selection bias toward continuous variables.
• Causal forests fitted only on usage data produce similar estimates to those using both usage and survey data, suggesting usage data captures much of the relevant heterogeneity.
• The distribution of individual treatment effects shows potential bimodality, indicating distinct groups of highly responsive and unresponsive households.
• Survey variables such as number of laptops, freezers, and employment status are among the most important demographic predictors, likely due to their correlation with income and energy use.